TL;DR: We propose Intrinsic Reward Policy Optimization (IRPO), a novel framework leveraging a surrogate policy gradient to overcome credit assignment and sample inefficiency in sparse-reward environments.
TL;DR: We propose a contraction actor-critic (CAC) algorithm for endowing a stability guarantee to the RL-trained policies for high-dimensional and nonlinear path-tracking problems.
TL;DR: We propose a macro-action discovery method and use it in a hierarchical algorithm for solving complex meta-RL problems.
TL;DR: We propose to use K-Mean clustering algorithm to measure the sparsity of each data point and use those measures to overestimate the probable safety violation for safe deployment of a policy.
TL;DR: We propose a policy optimization framework that provides adaptable safety guarantees on unseen tasks by viewing constrained reinforcement learning through the lens of meta-learning.
TL;DR: We propose to learn transition dynamics and reward function using causal normalizing flow model for out-of-distribution adaptation of a policy.
